Four Projects Have Made All Three Issues, and No Two Move Alike

From 2026‑09‑02 to 2026‑09‑09 the leaderboard is dominated by infra‑tooling repos that all tackle the same broad problem: giving AI coding agents a more disciplined, observable, or composable environment. Four names have now appeared on every issue we have published — Skills, Ponytail, DeepSeek Harness and Orca — and no two of them are moving the same way: one gains faster each time, one gains less each time, one has held its pace, and one swings with no direction at all. The rest of the list spreads across text‑LLM utilities, audio workstations and video rendering, showing that the community is still experimenting with many niches beyond core agent infrastructure.

#1 mattpocock/skills — a set of composable engineering skills for AI coding agents to replace vibe coding with rigorous software fundamentals

+13,010 stars this week → 256,951 · skills star curve

2026-W35: +14,911 · 2026-W36: +8,019 · up and down across 3 issues · Earlier: 213,421 stars on 2026-08-11

Skills supplies a library of tiny, prompt‑based instructions that let you inject classic software‑engineering practices—test‑driven development, domain modeling, disciplined debugging—directly into Claude Code, Codex or similar agents. By treating each practice as a composable “skill”, the repo lets you replace ad‑hoc, vibe‑driven code generation with a repeatable engineering workflow. Its gains have swung up and down rather than settling anywhere, so there is no trend here to read yet.

#2 tt-a1i/archify — a deterministic rendering system that turns AI-generated JSON into interactive system maps

+12,853 stars this week → 54,679 · archify star curve

2026-W36: +19,312 · down from the last issue · Earlier: 13,033 stars on 2026-08-16

Archify offers a Node.js rendering and validation pipeline that turns AI‑generated JSON descriptions of a system into interactive, verified maps. The idea is to give agents a deterministic way to visualise codebases or architectural specs instead of the usual static diagrams. It gained less this issue than last, its first cooling since the surge that took it to first place — though one step down is not yet a direction.

#3 DietrichGebert/ponytail — a simplicity-enforcing plugin for AI agents that eliminates over-engineering and code bloat

+12,233 stars this week → 132,166 · ponytail star curve

2026-W35: +5,000 · 2026-W36: +8,369 · gaining faster each issue, 3 issues running · Earlier: 100,614 stars on 2026-08-11

Ponytail is a plugin that forces AI agents to follow a strict ladder of simplicity before emitting any new code, preventing over‑engineering, boilerplate bloat and needless token consumption. The ladder checks for existing functionality, standard‑library support, native platform features and even one‑line feasibility before allowing a write. Unlike its infra‑tooling peers, Ponytail’s star gain is accelerating, indicating rising interest in lightweight, cost‑aware agent constraints.

#4 affaan-m/ECC — an agent harness operating system that provides a coordinated engineering system and toolbox for AI agents

+8,532 stars this week → 254,283 · ECC star curve

Earlier: 239,390 stars on 2026-08-11

ECC acts as an operating system for AI agents, stitching together planning, testing, implementation, review and verification steps into a coordinated engineering workflow. By providing a massive library of specialised agents, skills and runtime hooks, it turns a raw code‑writing model into a disciplined software‑engineering assistant.

#5 deepseek-ai/deepseek-harness — an open-source agent harness featuring an everything-is-a-plugin architecture

+8,162 stars this week → 216,221 · deepseek-harness star curve

2026-W35: +42,534 · 2026-W36: +16,749 · gaining less each issue, 3 issues running · Earlier: 83,753 stars on 2026-08-14

DeepSeek Harness supplies an open‑source, plugin‑centric framework for building and managing AI agents. Its “everything‑is‑a‑plugin” architecture lets developers compose capabilities in a spatiotemporal, JavaScript‑driven fashion, making the harness especially attractive for Node.js‑centric projects. It has gained less in each issue since it led the ranking in August, the steepest cooling on the board and the counterweight to Ponytail.

#6 debpalash/VoiceStudio — a local-first audio workstation for voice cloning, dubbing, and transcription

+7,528 stars this week → 21,317 · VoiceStudio star curve

Earlier: 9,969 stars on 2026-08-15

VoiceStudio is a local‑first desktop workstation that aggregates more than two dozen AI speech models behind a unified Tauri/React UI, letting you clone voices, dub media and transcribe audio without ever sending data to the cloud.

#7 blader/humanizer — a pattern-based text rewriter that strips AI-sounding clichés to create natural human-like prose

+5,771 stars this week → 45,401 · humanizer star curve

Earlier: 35,845 stars on 2026-08-16

Humanizer rewrites AI‑generated prose to sound more human by applying thirty‑five patterns derived from Wikipedia’s “Signs of AI writing”. It preserves factual content while stripping away the formulaic phrasing that makes text feel synthetic.

#8 stablyai/orca — an AI orchestrator for running multiple CLI coding agents in parallel isolated worktrees

+5,082 stars this week → 64,310 · orca star curve

2026-W35: +5,420 · 2026-W36: +5,522 · holding the same pace, 3 issues running · Earlier: 42,650 stars on 2026-08-11

Orca orchestrates multiple CLI‑based coding agents in parallel isolated Git worktrees, letting you fan out a single prompt, compare outputs side‑by‑side and merge the best results. Its star growth has held the same pace issue after issue, which reads as steady niche appeal rather than a spike.

#9 lnkiai/m3e-canvas — a visual sketching tool that converts Material 3 Expressive designs into AI coding prompts

+4,656 stars over 4.8 days → 5,206 · m3e-canvas star curve

M3E Canvas is a browser‑based design tool that lets you sketch Material 3 Expressive screens and instantly turn those visual layouts into natural‑language prompts for AI coding assistants, bridging design and code generation.

#10 heygen-com/hyperframes — a framework for rendering deterministic MP4 videos from HTML and CSS built for AI agents

+4,181 stars this week → 47,738 · hyperframes star curve

Earlier: 41,083 stars on 2026-08-15

HyperFrames lets developers write HTML, CSS and seekable animation scripts that are compiled into deterministic MP4 videos, a workflow that aligns well with AI agents tasked with automated video production.

Breaking out

Ranked by growth relative to their own size, not by stars added. These are small enough that a week can multiply them — and small enough to have escaped you.

Player-YN/PawWork_ZhuaZhua · +405.0% this week → 1,924 stars · PawWork_ZhuaZhua star curve

Earlier: 381 stars on 2026-09-02

PawWork tackles the tedious task of turning arbitrary web snippets into fully editable office‑style documents by letting users select content on a page and feed it to a BYOK AI model, which then produces spreadsheets, slide decks or HTML sites ready for further editing.

anthropics/commerce-agents · +345.7% over 5.8 days → 2,527 stars · commerce-agents star curve

Commerce‑Agents solves the problem of building safe, end‑to‑end AI shopping experiences by offering two ready‑made agents—one for customer‑facing browsing and checkout, the other for merchant‑side inventory and marketing—each wired into a secure backend.

rakanki911/DLSS5-Swapper · +344.6% over 5.8 days → 3,032 stars · DLSS5-Swapper star curve

DLSS5‑Swapper simplifies the otherwise manual process of adding neural rendering to games, presenting a unified UI that detects compatible titles and lets users toggle between native DLSS, Feeder‑based upscaling or other renderers with a single click.

experientiallabs/experiential · +279.5% this week → 3,093 stars · experiential star curve

Earlier: 353 stars on 2026-08-23

Experiential addresses the fragmentation of LLM usage in agent pipelines by exposing a single OpenAI‑compatible endpoint that aggregates providers, then leverages collected OpenTelemetry traces to automatically route traffic to the most efficient model.

zvec-ai/zvec-grep · +267.4% over 5.8 days → 3,207 stars · zvec-grep star curve

zvec‑grep unifies three disparate search paradigms—regex‑based ripgrep, BM25 lexical ranking and semantic vector similarity—into one command, giving developers and agents a flexible tool for locating precise or conceptually related information across large local workspaces.

The repeat entries are the argument of this issue: developers keep reaching for tools that constrain agents rather than tools that prompt them, but that appetite is not one thing moving together. Ponytail’s climb says the market for enforced simplicity is still opening, while DeepSeek Harness’s cooling and Orca’s unchanged pace suggest the full harness has found its level.